Safety Evaluation Policy

Defines mandatory safety evaluation procedures prior to model release.

All models and agentic systems prior to deployment.

Safety Evaluation.

Evaluations cover harmful content, bias/fairness, cybersecurity uplift, biological/chemical uplift, autonomous replication risk, deception/manipulation, and jailbreak resistance.

⟳Mandatory CBRN Expert Gate Any model demonstrating measurable uplift on biological, chemical, radiological, or nuclear benchmark categories automatically triggers mandatory external domain-expert review before any release decision, regardless of overall risk-level classification. This gate cannot be waived by internal sign-off alone.

⟳Multimodal Evaluation Evaluation suites now include cross-modal jailbreak resistance testing (image/audio/video-embedded instruction injection) for all multimodal-capable models.

Results documented in a Safety Report per the Safety Report Template, reviewed by AI Safety Research Lead before gate advancement.

AI Safety Research Lead owns evaluation design; Research Engineering executes; Governance Review Board reviews Level 3-4 results; external CBRN experts engaged per the mandatory gate.

Results retained minimum 5 years for audit.

Models failing thresholds blocked until remediated.

safety@webspaceai.in

Reduced-scope evaluation for minor updates requires AI Safety Research Lead sign-off; CBRN gate has no exceptions.

Quarterly, updated as new risk categories emerge.

safety@webspaceai.in